Loading video...

Video Failed to Load

Go Home

Reviewing AI-generated code is the new bottleneck. Writing code was slow before. Now, in a few minutes, you can generate virtually unlimited lines of code using AI. But developers are now spending 70% of their time checking that code. That's the new bottleneck. I've been using Cline, and they...

64,477 views • 11 months ago •via X (Twitter)

34 Comments

Santiago's profile picture
Santiago11 months ago

By the way, Cline was recently the fastest-growing AI project on GitHub, and it's fully open source with 50,000+ stars. Here is a link to install Cline on @code: If you want to understand the new "Explain Changes" feature, here is a link to their blog:

Kirha's profile picture
Kirha11 months ago

AI moved the bottleneck from writing code to trusting it. Tools that make verification faster are now more valuable than tools that make generation faster, because output is no longer the scarce resource, developer attention is Anything that compresses the “review + understand” loop is going to be essential in modern workflows

Bill's profile picture
Bill11 months ago

This isn't about saving time. It's because people generating all this code don't understand it, so they can't possibly review it. And somehow they think reviewing it in English is good enough.

Nick's profile picture
Nick11 months ago

@sdrzn Santiago found /explain-changes lfg

Vayo's profile picture
Vayo11 months ago

time to install cline...

Ajit Gupta's profile picture
Ajit Gupta11 months ago

New job in upcoming years Software developer ❌️ Software reviewer ✅️ Code is there, but we need to review it. So, you still need to know how to write code before reviewing AI's code.

Nikos Katsikanis's profile picture
Nikos Katsikanis11 months ago

if you can't review the code, then plumbing is a better move

juanmacias 🏳️‍🌈's profile picture
juanmacias 🏳️‍🌈11 months ago

You can put the AI to review the code, is what I’m doing…

Finleysson's profile picture
Finleysson11 months ago

Let another AI check it. Problem solved

andy's profile picture
andy11 months ago

we just externalized the 'typing fast' part and internalized the 'thinking hard' part. sounds like the bottleneck just upgraded its address, not disappeared.

Dan Wells's profile picture
Dan Wells11 months ago

Why bother? Why not just let it turn out crappy products & tell us we have to buy them & like them or else? That’s been the MO of tech for a while.

Ehsan 🇮🇷's profile picture
Ehsan 🇮🇷11 months ago

Nice, now instead of reading 100 lines of code, devs have to read 100 lines of text

Anirudh Sharma's profile picture
Anirudh Sharma11 months ago

What are your thoughts on Claude Code's "review" and "security-review" modes?

Milla Millennial's profile picture
Milla Millennial11 months ago

Unpopular opinion. If a reviewer needs an AI to explain the code they're supposed to review, then maybe they're not qualified enough.

Zaf____'s profile picture
Zaf____11 months ago

Have tried multiple prompts like this. The best way to review is not to look at the lines but to look at the impact. Tell the LLM to find the flows affected by this change and then show the before/after of the flow in Psuedo code. That's all.

Daniel King's profile picture
Daniel King11 months ago

Part of the reason why I built KarpeSlop, detecting AI slop was taking too much time

Brent Brookler's profile picture
Brent Brookler11 months ago

wait! cline or anything ? i’m confused.

John Au-Yeung's profile picture
John Au-Yeung11 months ago

This is such a smart approach. Code generation is fast, but understanding it safely is the real limiter. Pairing every line with an explanation is like adding automated pair programming. Even a mini-guide: review diffs in chunks with context notes to halve the mental load.

Prometheus Protocol's profile picture
Prometheus Protocol11 months ago

This is exactly the right observation. The bottleneck has shifted from creation to verification. It's the human-level version of the exact problem AI agents face at a protocol level. An agent can’t ask for a line-by-line explanation before using a tool; it needs a mathematical guarantee. That's why we built on-chain rails for a provable software supply chain, verifying every tool from git commit to on-chain deployment.

Massoud's profile picture
Massoud11 months ago

The output models need to improve. Solid code should be based on data structure objects and reliable well understood algorithms.

Mike Boysen (JTBD/acc)'s profile picture
Mike Boysen (JTBD/acc)11 months ago

That's not a bottleneck, it's a flufferbot. Focus on the real problem and resolve it. It won't be about people. Trust me

Nate Nguyen's profile picture
Nate Nguyen11 months ago

Code review taking 70% of the time? Ouch! 😅 Sounds like my last sprint... Cline looks promising. Anyone else tried it?

Sree Teja Laveti's profile picture
Sree Teja Laveti11 months ago

I feel like this is a new career in the upcoming phases of spare engineering! Soon there will be openings for code debuggers on AI written code!

ZeroLikes's profile picture
ZeroLikes11 months ago

I'm successfully vibe coding using the free version of chatgpt, literally nothing else is needed, just a good coder to begin with and the correct prompts.

C's profile picture
C11 months ago

For me it works fine, to let AI create a lot of mermaidjs chart (text2charts).. People (specially all the non coders involved in DEV workflows) could much faster understand and compare different solutions/changes

Sin Jeong-hun's profile picture
Sin Jeong-hun11 months ago

Also, the AI code generation is slow. I have to spend a lot of time waiting after prompting. If the generation was almost instant, reviewing would not be much of a problem for me.

C's profile picture
C11 months ago

What?? Review?? Just vibe push it, no? 😅

RedEyed's profile picture
RedEyed11 months ago

It's not because code is complex, it is because ai generated code is out of project context and low quality

Subcoder's profile picture
Subcoder11 months ago

lol sounds annoying... there's better ways to dev with AI.

Viraj Kawa 🇮🇳's profile picture
Viraj Kawa 🇮🇳11 months ago

This video alone does more explaining than most landing pages. Inline reasoning per line is exactly what AI code review needed.

Jerry Fernholz's profile picture
Jerry Fernholz11 months ago

git reset --hard The most important git command when working with AI coding agents 🤣.

FastFred68's profile picture
FastFred6811 months ago

It doesn't always work right.

Vadym Petryshyn's profile picture
Vadym Petryshyn11 months ago

So we automated the easy part and made the hard part 3x bigger. Classic engineering move. Now we need AI to review AI code, and then AI to review that review...

Ali Sherief's profile picture
Ali Sherief11 months ago

This could completely change how we onboard onto unfamiliar codebases. Context + explanation in one view is huge

Related Videos

👀 I used OpenAI's Code Interpreter to make Flappy Bird 🐦in 7 minutes: Code Interpreter/GPT-4 for code generation. Pre-existing or AI-generated assets for graphics. --- Here's how to make the game in only 6 steps: (1): Enter the following prompt: "write p5.js code for Flappy Bird where you control a yellow bird continuously flying between a series of green pipes. The bird flaps every time you left click the mouse. If the bird falls to the ground or hits a pipe, you lose. This game goes on infinitely until you lose and you get points the further you go". (2): Use generative AI or existing game assets and spirits. I searched "flappy bird assets" on Google and used the first link, a GitHub repo with pngs from the original Flappy Bird. (3): Use this prompt to link assets to the code: "Please generate the entire file again based on the fact I'm using a unique background, spirits for the bird, and pipes. Here is the list of assets I'm using: [list of file names]." Code Interpreter should modify the code accordingly to include the list of file names. (4) Make an account OpenProcessing -> create a sketch -> paste in the code generated by Code Interpreter -> upload in-game assets from step (2). (5) (Optional) Ask ChatGPT to make changes to improve the in-game experience e.g., adding a high score, restarting the game when the bird dies, etc. Copy the new code into your OpenProcessing sketch and reload the game. (6) If something doesn't work, ask GPT4 to fix it. Copy and paste the error message and ask it to regenerate the code. --- Bonus Tips: - Iteratively test code. Each time you make a change using Code Interpreter, test the updated code by playing the game so you catch new bugs early. - Learn programming by asking questions: "Act as a senior programmer very good at explaining concepts to a beginner. Tell me how gravity works in this game and how you used code to make this happen."Code Interpreter/GPT4 for code generation. Download Pre-existing assets or generate new images for graphics. Excited to see what you make!

Alex Ker 🔭

739,874 views • 3 years ago